Work

The warehouse ran on a React Native app and web tools that 120+ staff relied on every shift, and they'd grown inconsistent and hard to change safely.

Once I'd spent enough time on the warehouse floor, a bigger problem showed up. Part-used packs of medication were being thrown away, and that waste added up to a lot of money.

I owned the React Native warehouse app and led an 18-month rebuild of the operational tools, one workflow at a time. I introduced testing infrastructure and a shared design system, and rolled each change out with the warehouse team before full release.

I pitched medication recovery, worked with the data team to target the most expensive medicines first, and shipped it in eight weeks as the product engineer alongside a PM, a designer and a backend engineer. The app walks staff through eligibility, prep, collection and return to stock. Anything that fails a check stops rather than guessing, and batch and expiry data stay attached so a recall can still find the stock.

Measured savings of about £265,000 a month. The rebuild ran with zero downtime for 120+ daily users, in a business serving millions of customers.

Brand deals arrive as a mix of PDFs, emails and screenshots of DMs. Creators end up copying fees, deliverables and posting dates into spreadsheets by hand, and details get missed.

A drop-in flow: upload whatever the brand sent and the app reads it into one structured deal, which the creator checks beside the originals before anything goes to the client. Campaigns, clients and amounts owed all live in one place, on desktop and mobile.

For the extraction I built a synthetic test set of good and broken inputs, checked required fields in code, and used a second model to judge whether the content was right. Then I compared models on that set and picked the one that did best.

Session recordings showed where new creators were dropping off, so I rebuilt onboarding around the end goal: drop in a brief, get a completed deal, then a guided tour of it.

Creators go from a messy inbox to a confirmed deal in one flow, and every screen is built from one set of shared components so the app stays consistent as it grows.

I kept needing the same skills to write SEO content with AI. A bare prompt gets you something generic, and the tools that do better run their own models and charge per article.

A hosted MCP endpoint your own agent calls. It stores a profile of the business, pulls keyword and Google Search Console data, hands the agent a writing brief, and runs deterministic quality gates that block a draft from publishing if it fails.

Around it sits a full SaaS: sign-up and onboarding, teams, Stripe billing, and a dashboard for articles and projects. I own all of it, from the schema and row-level security policies to Playwright end-to-end tests covering auth, teams, payments and the MCP flow.

Live with paying users. Built and run solo since June 2026, with about 560 commits and 38 migrations so far.

Every NHS trust publishes board papers, often over a hundred pages each, with budget signals and procurement plans buried inside. There are more than 300 trusts, each with its own website, so suppliers were reading them one PDF at a time.

The whole platform: sign-up, team accounts, and a background pipeline that sends each paper to Google Gemini along with the supplier's own company context and the themes they care about. It comes back as structured data: an opportunity score, the relevant sections, key people and a drafted sales pitch.

Every insight links to the exact page and section, so a buyer can check it in the original before using it. Around that sat Stripe billing with usage limits per plan, and a marketing site with generated pages for each trust and region.

Tested with real customers and led to good conversations. The market turned out too niche, with too much compliance overhead for a self-serve product, so I wound it down. It solved the problem well enough that it led to direct work with one of those businesses.

Patients recovering at home take readings with Bluetooth medical devices, and most of them are aged 60 to 90. Readings have to arrive reliably even when the connection doesn't behave, in a regulated medical-device setting.

Separately, engineers kept interrupting other teams to ask whether a noisy alert mattered.

I owned patient-facing features from requirement to release: integrating devices through their SDKs and protocols, designing the API layers to outside services, and working with external engineering teams to get them live. I ran on-site research with patients and staff and led the rebrand of the app. It looks deliberately plain: big type, one thing per screen, nothing clever.

I pitched an alert-triage bot to the CTO and built the MVP. It watches the triage channels in Slack and explains, with references, whether an alert is known noise or a real warning and what the service does.

The app is live with Doccla's patients. The bot reached MVP before I left, so I can't point to measured savings from it.

Operations data was spread across several sources, with no single live view of what was happening.

A real-time dashboard in Next.js, backed by a single GraphQL API that unified the data sources behind it. I also built a greenfield design system and component library for it.

Teams got one live view of operations instead of several. This was an employer project, so I can describe it but not share screens.